Identification of soil profile classes using depth-weighted visible–near-infrared spectral reflectance

Abstract

Soil visible–near-infrared (Vis-NIR) diffuse reflectance, which can be collected rapidly and cheaply, contains rich information of the properties that are useful for soil taxonomic diagnosis. Vis-NIR spectroscopy has the potential to identify soil taxonomic classes in an efficient and cost-effective way. This study explores the potential of Vis-NIR spectroscopy to identify soil profile classes at the order, suborder, group, and subgroup levels of the Chinese Soil Taxonomy (CST).The authors measured the Vis-NIR (350–2500 nm) diffuse reflectance of 2260 legacy soil samples, which were sampled by genetic horizon and collected from 527 soil profiles (75% for training and 25% for validation). To represent the overall spectral pattern for each soil profile, the depth-weighted average was used to combine the spectral reflectance of genetic horizons. The synthetic minority over-sampling technique (SMOTE) was used to obtain balanced training data. Principal component analysis (PCA) was performed to extract spectral predictors used for random forest (RF) modeling in soil classification.Vis–NIR spectral reflectance exhibited acceptable overall performance for identification of soil orders and suborders, with overall validation accuracies of 0.63 and 0.62, respectively, but low overall performance at group and subgroup levels, with overall validation accuracies of 0.40 and 0.28, respectively. The overall performance at different taxonomic levels was affected by the number of soil classes and the class distribution of the soil profiles. Soil classes with pedogenic processes associated closely with spectrally active soil properties or with characteristic profile patterns were most accurately identified, even if they were minority classes. The results show that the Vis–NIR spectral pattern of soil profile can be used to identify soil profile classes at higher taxonomic levels in the CST system. Combined with machine-learning techniques, the soil Vis–NIR spectral library will serve as an efficient tool for digital soil survey mapping and updating with the use of legacy soil samples and the reduction of conventional laboratory analyses.